#include <mach/ppc/asm.h>
#undef ASSEMBLER
+#define __APPLE_API_PRIVATE
+#include <machine/cpu_capabilities.h>
+#undef __APPLE_API_PRIVATE
+
+/* We use mode-independent "g" opcodes such as "srgi". These expand
+ * into word operations when targeting __ppc__, and into doubleword
+ * operations when targeting __ppc64__.
+ */
+#include <architecture/ppc/mode_independent_asm.h>
+
+
// ***************
// * S T R C A T *
// ***************
// We optimize the move by doing it word parallel. This introduces
// a complication: if we blindly did word load/stores until finding
// a 0, we might get a spurious page fault by touching bytes past it.
-// To avoid this, we never do a "lwz" that crosses a page boundary,
+// To avoid this, we never do a load that crosses a page boundary,
// and never store a byte we don't have to.
//
// The test for 0s relies on the following inobvious but very efficient
// The test maps any non-zero byte to zero, and any zero byte to 0x80,
// with one exception: 0x01 bytes preceeding the first zero are also
// mapped to 0x80.
+//
+// In 64-bit mode, this algorithm is doubleword parallel.
.text
.globl EXT(strcat)
.align 5
-LEXT(strcat) // char* strcat(const char *s, const char *append);
- andi. r0,r3,3 // is dst aligned?
- dcbtst 0,r3 // touch in dst
- lis r6,hi16(0xFEFEFEFF) // start to load magic constants
+LEXT(strcat) // char* strcat(const char *s, const char *append);
+ clrrgi r9,r3,LOG2_GPR_BYTES// align pointer by zeroing right LOG2_GPR_BYTES bits
+ li r10,-1 // get 0xFFs
+ lg r8,0(r9) // get word or doubleword with 1st operand byte
+ rlwinm r11,r3,3,(GPR_BYTES-1)*8 // get starting bit position of operand
+#if defined(__ppc__)
+ lis r6,hi16(0xFEFEFEFF) // start to generate 32-bit magic constants
lis r7,hi16(0x80808080)
- dcbt 0,r4 // touch in source
+ srw r10,r10,r11 // create a mask of 0xFF bytes for operand in r8
ori r6,r6,lo16(0xFEFEFEFF)
ori r7,r7,lo16(0x80808080)
- mr r9,r3 // use r9 for dest ptr (must return r3 intact)
- beq Lword0loop // dest is aligned
- subfic r0,r0,4 // r0 <- #bytes to word align dest
- mtctr r0
-
-// Loop over bytes looking for 0-byte marking end of dest.
-// r4 = source ptr (unalaigned)
-// r6 = 0xFEFEFEFF
-// r7 = 0x80808080
-// r9 = dest ptr (unaligned)
-// ctr = byte count
-
-Lbyte0loop:
- lbz r8,0(r9) // r8 <- next dest byte
- addi r9,r9,1
- cmpwi r8,0 // test for 0
- bdnzf eq,Lbyte0loop // loop until (ctr==0) | (r8==0)
-
- bne Lword0loop // enter word loop if we haven't found the 0-byte
- subi r9,r9,1 // point to 0-byte
- b L0found // start to append the source
-
-// Loop over words looking for 0-byte marking end of dest.
-// r4 = source ptr (unalaigned)
+#else
+ ld r6,_COMM_PAGE_MAGIC_FE(0) // get 0xFEFEFEFE FEFEFEFF from commpage
+ ld r7,_COMM_PAGE_MAGIC_80(0) // get 0x80808080 80808080 from commpage
+ srd r10,r10,r11 // create a mask of 0xFF bytes for operand in r8
+#endif
+ orc r8,r8,r10 // make sure bytes preceeding operand are nonzero
+ b Lword0loopEnter
+
+// Loop over words or doublewords looking for 0-byte marking end of dest.
+// r4 = source ptr (unaligned)
// r6 = 0xFEFEFEFF
// r7 = 0x80808080
-// r9 = dest ptr (word aligned)
+// r9 = dest ptr (aligned)
.align 5 // align inner loops for speed
Lword0loop:
- lwz r8,0(r9) // r8 <- next dest word
- addi r9,r9,4
+ lgu r8,GPR_BYTES(r9) // r8 <- next dest word or doubleword
+Lword0loopEnter: // initial entry
add r10,r8,r6 // r10 <- word + 0xFEFEFEFF
andc r12,r7,r8 // r12 <- ~word & 0x80808080
and. r11,r10,r12 // r11 <- nonzero iff word has a 0-byte
beq Lword0loop // loop until 0 found
-
- slwi r0,r8,7 // move 0x01 bits (false hits) into 0x80 position
- subi r9,r9,4 // back r9 up to beginning of word
- andc r11,r11,r0 // mask out false hits
- cntlzw r0,r11 // find 0 byte (r0 = 0, 8, 16, or 24)
- srwi r0,r0,3 // now r0 = 0, 1, 2, or 3
- add r9,r9,r0 // now r9 points to the 0-byte in dest
-
-// End of dest found, so we can start appending source.
-// We align the _source_, which allows us to avoid all worries about
-// spurious page faults. Doing so is faster than aligning the dest.
+
+// Now we know one of the bytes in r8 is zero, we just have to figure out which one.
+// We have mapped 0 bytes to 0x80, and nonzero bytes to 0x00, with one exception:
+// 0x01 bytes preceeding the first zero are also mapped to 0x80. So we have to mask
+// out the 0x80s caused by 0x01s before searching for the 0x80 byte. Once the 0 is
+// found, we can start appending source. We align the source, which allows us to
+// avoid worrying about spurious page faults.
// r4 = source ptr (unaligned)
// r6 = 0xFEFEFEFF
// r7 = 0x80808080
-// r9 = ptr to 0-byte (unaligned)
+// r8 = word or doubleword with a 0-byte
+// r9 = ptr to the word or doubleword in r8 (aligned)
+// r11 = mapped word or doubleword
-L0found:
- andi. r0,r4,3 // is source aligned?
- beq LwordloopEnter // skip if so
- subfic r0,r0,4 // not aligned, get #bytes to align r4
+ slgi r10,r8,7 // move 0x01 bits (false hits) into 0x80 position
+ andi. r0,r4,GPR_BYTES-1 // is source aligned?
+ andc r11,r11,r10 // mask out false hits
+ cntlzg r10,r11 // find 0 byte (r0 = 0, 8, 16, or 24)
+ subfic r0,r0,GPR_BYTES // get #bytes to align r4
+ srwi r10,r10,3 // now r0 = 0, 1, 2, or 3
+ add r9,r9,r10 // now r9 points to the 0-byte in dest
+ beq LwordloopEnter // skip if source is already aligned
+
mtctr r0 // set up loop
// Loop over bytes.
bne LwordloopEnter // 0-byte not found, so enter word loop
blr // 0-byte found, done
-// Word loop: move a word at a time until 0-byte found.
-// r4 = source ptr (word aligned)
+// Word loop: move a word or doubleword at a time until 0-byte found.
+// r4 = source ptr (aligned)
// r6 = 0xFEFEFEFF
// r7 = 0x80808080
// r9 = dest ptr (unaligned)
.align 5 // align inner loop, which is 8 words ling
Lwordloop:
- stw r8,0(r9) // pack word into destination
- addi r9,r9,4
+ stg r8,0(r9) // pack word or doubleword into destination
+ addi r9,r9,GPR_BYTES
LwordloopEnter:
- lwz r8,0(r4) // r8 <- next 4 source bytes
- addi r4,r4,4
+ lg r8,0(r4) // r8 <- next 4 or 8 source bytes
+ addi r4,r4,GPR_BYTES
add r10,r8,r6 // r10 <- word + 0xFEFEFEFF
andc r12,r7,r8 // r12 <- ~word & 0x80808080
and. r0,r10,r12 // r0 <- nonzero iff word has a 0-byte
- beq Lwordloop // loop if ctr!=0 and cr0_eq
+ beq Lwordloop // loop if no 0-byte
// Found a 0-byte. Store last word up to and including the 0, a byte at a time.
-// r8 = last word, known to have a 0-byte
+// r8 = last word or doubleword, known to have a 0-byte
// r9 = dest ptr
Lstorelastbytes:
- srwi. r0,r8,24 // right justify next byte and test for 0
- slwi r8,r8,8 // shift next byte into position
+ srgi. r0,r8,GPR_BYTES*8-8 // shift leftmost byte into bottom so we can "stb"
+ slgi r8,r8,8 // move on to next
stb r0,0(r9) // pack into dest
addi r9,r9,1
bne Lstorelastbytes // loop until 0 stored
projects / apple / libc.git / blobdiff